FOR
ALL-OPTICAL
WAVELENGTH DIVISION MULTIPLEXED
COMPUTER NETWORKS
By
Krishna Moorthy Sivalingam
June 1994
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF COMPUTER SCIENCE AND THE FACULTY OF THE GRADUATE SCHOOL OF THE STATE UNIVERSITY OF NEW YORK AT BUFFALO
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF
by
Krishna Moorthy Sivalingam
All Rights Reserved
Optic fiber is rapidly gaining significance as the primary medium in computer and telecom-munication networks. The principal advantages of optic fiber over copper cable are its higher transmission speed and lower error rates. In addition, parallel multiple channels are made available on a single fiber using a technique called Wavelength Division Multiplexing (WDM). This dis-sertation studies the design, analysis and implementation of communication protocols which fully exploit the high speed, low error rate characteristics and the availability of multiple channels on a single fiber. The protocols studied are used to support interprocessor communication in an optically interconnected multiprocessor system.
The first part of the dissertation studies media access protocols for a passive star coupled WDM network. The star topology was shown to have advantages over the bus topology with respect to larger system size and better fault tolerance. The protocols studied are based on an architecture where each node has a single tunable transmitter and a fixed receiver. It evaluates the advantages of these schemes with respect to performance, cost and implementation.
Photonic amplification significantly increases the number of nodes that can be connected to the network. This motivates reconsideration of bus based optical networks. The second part of the dissertation studies the performance of the bus and star topologies with amplification with respect to system size, protocol performance, fault tolerance, cost, and implementation.
A scalable, hierarchical, all-optical multiprocessor architecture that supports distributed shared memory (DSM) has been considered earlier. The performance of the associated cache coherence protocols and the DSM organization is largely dependent on media access protocol performance. The third part of the dissertation examines protocols for this environment. The protocols studied are optimized for the bi-modal traffic characteristics of a DSM system. Reservation based protocols are compared to static allocation protocols.
Analytical modeling techniques based on stochastic processes and semi-Markov models are used to evaluate protocol performance. All the analytic models are validated through discrete-event simulation.
I thank my adviser Dr. Patrick Dowd for his incessant support, encouragement, guidance, and advice that made this dissertation possible. His emphasis on the quality of research and writing have been extremely valuable in writing this dissertation.
I thank Dr. Sreejit Chakravarty and Dr. Wennie Shu for accepting to be on my committee and sparing precious time for discussions. Their comments, involvement and support have played a significant role in this research. I also thank Prof. Aura Ganz of University of Massachusetts, Amherst, for suggesting changes that assisted in improving this dissertation. The following acknowledges research done jointly with other members of the group: The research on pre-allocation media access protocols for WDM networks was developed with Dr. Kalyani Bogineni. The protocol synchronization mechanisms for a hierarchical network architecture were developed jointly with Dan Crouse. I would like to thank all members of Dr. Dowd’s research group with whom I have had frequent discussions: Khaled Aly, Eric Blade, Kalyani Bogineni, John Chu, Dan Crouse, David Hoffmeister, and Jim Perreault.
My gratitude to my parents for their infinite patience and implicit faith in my capabilities is boundless and cannot be expressed in sufficient words. My thanks to my brother Prabhu are limitless for consistently invigorating me to remain focused on my research. This dissertation is dedicated to them. Special mention goes to my uncle Dr. Kandaswami and aunt Mrs. Kandaswami for taking exceptional care of me throughout my studies. I take this opportunity to thank K. Sivakumar and Ajay Shekhawat who have rendered help beyond measure. Last but not the least, I am extremely grateful to Thriveni for her unceasing cheer and steadfast optimism.
1 Introduction 1
1.1 Fiber optic networks
: : : : : : : : : : : : : : : : : : : : : : : : : : : :
1 1.2 Access Protocols for Star: : : : : : : : : : : : : : : : : : : : : : : : :
3 1.3 Access Protocols for Hierarchical Architecture: : : : : : : : : : : : : :
42 Photonic Networks 7 2.1 OSI Model
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
9 2.2 WDM Networks: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
12 2.2.1 Optical Components: : : : : : : : : : : : : : : : : : : : : : : :
12 2.2.2 Network Topology: : : : : : : : : : : : : : : : : : : : : : : : :
17 2.3 WDM Media Access: : : : : : : : : : : : : : : : : : : : : : : : : : : :
20 2.3.1 Network Classification: : : : : : : : : : : : : : : : : : : : : : :
20 2.3.2 Single Channel Multiple Access: : : : : : : : : : : : : : : : : :
23 2.3.3 WDM Multiple Access: : : : : : : : : : : : : : : : : : : : : :
24 2.3.4 Node Architecture: : : : : : : : : : : : : : : : : : : : : : : : :
26 2.4 Summary: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
293 Access Protocols for WDM Star Networks 30
3.2.1 Random Access Protocols
: : : : : : : : : : : : : : : : : : : : :
36 3.2.2 Static Access Protocols: : : : : : : : : : : : : : : : : : : : : :
40 3.3 Performance Analysis: : : : : : : : : : : : : : : : : : : : : : : : : : :
44 3.3.1 Semi-markov model for I-SA: : : : : : : : : : : : : : : : : : :
46 3.3.2 Semi-Markov model for I-TDMA: : : : : : : : : : : : : : : : :
50 3.3.3 Computation of Performance Metrics: : : : : : : : : : : : : : :
55 3.3.4 Validation of Analytic Models: : : : : : : : : : : : : : : : : : :
58 3.3.5 Comparison of I-SA and I-TDMA: : : : : : : : : : : : : : : : :
61 3.3.6 Comparison of I-SA and I-TDMA*: : : : : : : : : : : : : : : :
63 3.3.7 Comparison of I-TDMA and I-TDMA*: : : : : : : : : : : : : :
69 3.3.8 Comparison of I-SA* and I-TDMA*: : : : : : : : : : : : : : :
70 3.3.9 Comparison to reservation based protocols: : : : : : : : : : : :
73 3.4 Summary: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
764 Star and Bus Topologies 78
4.1 Topologies
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
79 4.2 Access Arbitration Mechanisms: : : : : : : : : : : : : : : : : : : : : :
83 4.2.1 Folded Bus: : : : : : : : : : : : : : : : : : : : : : : : : : : :
83 4.2.2 Dual Bus: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
85 4.2.3 Star: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
85 4.3 Performance Analysis: : : : : : : : : : : : : : : : : : : : : : : : : : :
86 4.3.1 Performance Under Uniform Reference Model: : : : : : : : : :
89 4.3.2 Performance under Client/Server Model: : : : : : : : : : : : : :
95 4.3.3 Discussion: : : : : : : : : : : : : : : : : : : : : : : : : : : : :
97 4.4 Summary: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
1025.1 Lightning Architecture
: : : : : : : : : : : : : : : : : : : : : : : : : : :
105 5.2 Parallel Computer Memory Organization: : : : : : : : : : : : : : : : :
109 5.2.1 Shared Memory: : : : : : : : : : : : : : : : : : : : : : : : : :
110 5.2.2 Distributed Memory: : : : : : : : : : : : : : : : : : : : : : : :
111 5.2.3 Distributed Shared Memory: : : : : : : : : : : : : : : : : : : :
111 5.3 Protocol Definition: : : : : : : : : : : : : : : : : : : : : : : : : : : : :
112 5.3.1 Single-Level Protocols: : : : : : : : : : : : : : : : : : : : : : :
114 5.3.2 Multi-Level protocols: : : : : : : : : : : : : : : : : : : : : : :
121 5.4 Analytic Performance Model and Validation: : : : : : : : : : : : : : : :
122 5.4.1 State Definitions: : : : : : : : : : : : : : : : : : : : : : : : : :
124 5.4.2 Steady state probabilities: : : : : : : : : : : : : : : : : : : : :
129 5.4.3 Performance metrics: : : : : : : : : : : : : : : : : : : : : : : :
132 5.4.4 Model Validation: : : : : : : : : : : : : : : : : : : : : : : : :
133 5.5 Performance Analysis: : : : : : : : : : : : : : : : : : : : : : : : : : :
136 5.5.1 Single-Level system performance: : : : : : : : : : : : : : : : :
136 5.5.2 Two-Level system performance: : : : : : : : : : : : : : : : : :
142 5.5.3 Slotted Aloha for control cycle: : : : : : : : : : : : : : : : : :
145 5.6 Synchronization: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
154 5.6.1 Bit Synchronization: : : : : : : : : : : : : : : : : : : : : : : :
155 5.6.2 Slot Synchronization: : : : : : : : : : : : : : : : : : : : : : : :
161 5.6.3 Frame Synchronization: : : : : : : : : : : : : : : : : : : : : :
165 5.7 Summary: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
170 6 Conclusions 1713.1 Notation used in analysis of the I-SA and I-TDMA.
: : : : : : : : : : : :
44 4.1 Topology comparison, whereM
is the number of nodes.: : : : : : : : :
81 4.2 Comparison of protocols to study impact of processing latency forC
=16,M
= 32,P
= 1, = 0:
5, and 2 f0;
1;
2;
4;
8g. The table shows the actual value of delay and throughput for=0 and the percentage variation for other values of.: : : : : : : : : : : : : : : : : : : : : : : : : : : :
95 4.3 Transceiver requirements of the protocols.: : : : : : : : : : : : : : : : :
98 4.4 Performance of protocols forC
= 16,M
= 32, = 0, = 0:
5, andP
2f1;
4;
8;
16g. The actual delay and throughput are presented forP
=1 and percentage increase/decrease for other values ofP
.: : : : : : : : : :
101 5.1 Notation used in analysis of FatMAC.: : : : : : : : : : : : : : : : : : :
130 5.2 State transition probabilities.: : : : : : : : : : : : : : : : : : : : : : : :
130 5.3 Expected number of packets in each state used in Little’s law to derive2.1 OSI model for computer communication.
: : : : : : : : : : : : : : : : :
10 2.2 Optical multiple access channel configurations: (a) folded unidirectionalbus, (b) doubly folded unidirectional bus, (c) dual unidirectional bus, and (d) star-coupled configuration.
: : : : : : : : : : : : : : : : : : : : : : :
18 2.3 Shufflenet implementation on the physical star topology.: : : : : : : : :
22 2.4 Performance characterization of Random and Static Access protocols.: :
24 2.5 Transmitter/receiver configurations for a WDM network.: : : : : : : : :
27 3.1 Network architecture based on the star topology with one tunable transmitterand fixed receiver per node.
: : : : : : : : : : : : : : : : : : : : : : : :
34 3.2 Allocation map for I-TDMA and I-TDMA* forC < M
and=0.: : : :
40 3.3 State diagram of the semi-markov process for analyzing the I-SA protocol. 47 3.4 State diagram of the semi-markov process for analyzing the I-TDMA protocol. 51 3.5 Validation of I-SA forM
2 f8;
16;
32g andC
2 fM;M=
2;M=
4g: (a)Average Packet Delay, (b) Network Throughput. Points and lines represent simulation and analysis respectively.
: : : : : : : : : : : : : : : : : : : :
59 3.6 Validation of I-TDMA forM
2f8;
16;
32gandC
2fM;M=
2;M=
4g: (a)Average Packet Delay, (b) Network Throughput. Points and lines represent simulation and analysis respectively.
: : : : : : : : : : : : : : : : : : : :
60M
32 andC
2;
4;
8;
16;
32 , (b)M
C
forC
8;
16;
32 .: : : :
62 3.8 Comparison of I-SA and I-TDMA* forB
= 100: (a)M
= 32 andC
2f1
;
2;
4;
8;
16;
32g, (b)C
=M
andM
2f8;
16;
32g.: : : : : : : : : : :
63 3.9 Impact of queue capacity: (a) I-SA forM
2 f16;
32;
64g,C
=M=
2 andB
2f2;
4;
8;
100g, (b) I-TDMA* forM
2f16;
32;
64g,C
=M=
2 andB
2 f2;
4;
8;
100g, (c) I-SA and I-TDMA* forM
=32,C
2fM=
4;M=
2;M
g and total queue capacity = 2C
, (d) Comparison of I-SA and I-TDMA* forM
2f16;
32;
64g,C
=M=
2 and total queue capacity = 2C
.: : : : : : :
64 3.10 Performance of I-TDMA* forM
=32 andC
2f8;
16;
32g: (a)2f0;
1g(b)
=4.: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
69 3.11 Comparison of I-SA* and I-TDMA* forM
2 f16;
32;
64g,C
=M=
2,I=
M=
2,A=M=
2: (a)=0, (b)=1, (c)=2, (d)=4.: : : : :
71 3.12 Comparison of I-SA and I-TDMA* toP6 – a reservation based protocol:(a)
M
= 32;C
2 f8;
16;
32gandL
= 32, (b)M
2 f32;
64g,C
=M=
8 andL
=32.: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
73 4.1 Allocation map forM
=8 nodes on one channel for B-TDMA. Internodepropagation delay is assumed to be =1 and cycle length is determined to be 9 slots.
: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
85 4.2 Performance of D-Net/C, Fairnet and B-TDMA forM
= 32, and varyingC
and propagation delay: (a) D-Net/C, (b) Fairnet, and (c) B-TDMA.: :
90 4.3 Performance of EQEB, I-SA* and I-TDMA* forM
=32 and varyingC
and propagation delay: (a) EQEB, (b) I-SA*, and (c) I-TDMA*.
: : : : :
93 4.4 Performance comparison of protocols with client server model forC
=16,5.2 Example of 24 node 3 level structure.
: : : : : : : : : : : : : : : : : : :
107 5.3 Parallel Computer memory organization: (a) shared memory, (b) distributedmemory, and (c) distributed shared memory.
: : : : : : : : : : : : : : : :
110 5.4 Assignment map for single level I-TDMA and FatMAC protocols forM
=6 and
C
=2: (a) I-TDMA – one slot equals memory block packet, and (b) FatMAC – one slot equals control packet.: : : : : : : : : : : : : : : : :
115 5.5 Assignment map for two level I-TDMA and FatMAC protocols forM
=44 and
C
=(2;
2): (a) I-TDMA, and (b) FatMAC.: : : : : : : : : : :
120 5.6 Semi-Markov analytic model for FatMAC protocol. Each transmitter hascapacity to hold one packet when processing another packet. Table 2 lists the transition probabilities.
: : : : : : : : : : : : : : : : : : : : : : : : :
123 5.7 Validation of analytic model for FatMAC protocol through simulation forM
2f16;
32g,C
2f2;
4;
8g,2 f0
:
5;
1:
0gandL
=16. Points and lines represent simulation and analysis respectively.: : : : : : : : : : : : : : :
134 5.8 Comparison of single-level protocol performance forM
= 16 andC
2f1
;
2;
4;
8g: (a) - (c)L
=32: (a)=0
:
1, (b)=0
:
5, (c)=0
:
9; (d) - (f)L
=64: (d)=0
:
1, (e)=0
:
5, (f)=0
:
9.: : : : : : : : : : : : : : :
137 5.9 Comparison of maximum throughput for I-TDMA and FatMAC forL
=128 with varying
and
C
: (a)M
=16, and (b)M
=64.: : : : : : : : :
139 5.10 Comparison of single-level protocol performance forM
2 f16;
32;
64g,(a)–(c)
M
=44,L
=32,C
2f(1;
3);
(2;
2);
(3;
1)g, (a)=0
:
1, (b)= 0
:
5, (c)= 0
:
9; (d)–(f)M
=44,L
=64,C
2 f(1;
3);
(2;
2);
(3;
1)g, (d)=0
:
1, (e)=0
:
5, (f)=0
:
9.: : : : : : : : : : : : : : : : : : :
143 5.12 Comparison of two-level protocol performance forM
= 44,L
= 64,C
2f(1;
3);
(2;
2);
(3;
1)g,p
1=0:
7 andp
2 =0:
3: (a)=0
:
5, (b)=0
:
9. 144 5.13 Comparison of performance with TDM and SA on control channel forC
= 4,L
= 172,= 0
:
3 at light loads forM
2 f16;
32;
64;
128g. Propagation delay is varied as 10m, 100m, 500m, 1Km, 2Km: (a) Speed of 200 Mbps, and (b) Speed of 2.56 Gbps.: : : : : : : : : : : : : : : : :
151 5.14 Comparison of SA and TDM on Control channel for varyingM
,C
= 4,speed of 2.56 Gbps, and
L
= 172. Propagation delay is varied as 10m, 100m, 500m, 1Km and 2Km at a load 50 packets per second per node.: :
154 5.15 Study ofmax forM
=16,L
2f32;
64g,2f0
:
1;
0:
25;
0:
5gand varyingC
: (a)L
=32, and (b)L
=64.: : : : : : : : : : : : : : : : : : : : : :
160 5.16 Simulation study of impact of propagation delay on lock-step and distributedclock algorithms for
M
=16,C
2f2;
4;
8g, andL
=64: (a)=0
:
1; (b)=0
:
5; and (c)=0
:
9.: : : : : : : : : : : : : : : : : : : : : : : : :
164 5.17 Simulation study of impact of propagation delay on frame synchronizationschemes for
M
=16,C
2f2;
4;
8g, andL
=64: (a)=0
:
1; (b)=0
:
5; and (c)=0
:
9.: : : : : : : : : : : : : : : : : : : : : : : : : : : : : :
168Introduction
This dissertation studies the design, analysis and implementation of high speed communi-cation protocols for optical networks. Networks based on a technique called wavelength division multiplexing that support multiple wavelength transmission on a single fiber are studied. These networks have many potential areas of application: local, metropolitan, and wide area networks, processor interconnection in multiprocessor systems, and heteroge-neous cluster-based computing systems. This dissertation emphasizes the study of media access protocols for an optically interconnected distributed shared memory multiprocessor system. The following sections introduce the basic concepts of the issues studied in the dissertation.
1.1
Fiber optic networks
Optical fiber is becoming widespread as the communication medium and is expected to be the dominant mode of transmission in the future. With the maturing of optical fiber technology, transmission costs and in particular cost of high speed data links has dropped tremendously in the last few years.
Fiber optic networks are characterized by very high transmission speeds and very low error rates. Conventional wide area networks based on twisted pair or coaxial cable operate at speeds of 64 Kbps to up to 45 Mbps. Local area networks such as Ethernet operate at 10 Mbps. Copper media can support transmission rates up to Gbps. On the other hand, the fiber has a potential bandwidth of the order of 25000 to 30000 Gbps [1]. Also, the fiber optic networks are characterized by orders of magnitude better error rate (bit error rate of 10?15
compared to 10?5
of previous networks) [2]. Other advantages of optic networks include smaller size, lighter weight, lower attenuation, electromagnetic isolation, and greater repeater spacing [2].
The electronic processing capability of the computers has not increased to meet the high speed of the communication medium. In conventional networks, electronic processing speed is faster than the transmission speed. In fiber optic networks, transmission speed is much higher than electronic processing speed which is typically limited to a few gigabits per second. The communications bottleneck has thus shifted from the transmission medium to the processing medium. This requires redesign of the network protocols to efficiently utilize the high speeds and low error rates.
One technique to minimize the speed mismatch between processing and transmission is Wavelength Division Multiplexing (WDM) [1]. Multiple optical channels can be formed on a single fiber using this technique, where data can be simultaneously transmitted on different wavelengths on the same fiber. This is achieved using tunable transmitters and/or receivers at each network node. Thus, it is possible to establish concurrent communication between more than one source-destination pair using a single fiber. The number of such channels is typically of the order of 20-30 channels [1], but the feasibility of having more than 100 channels has been demonstrated [3]. Each of these channels operate closer to the maximum speed of the electronic processing units which is of the order of Gigabits per
second.
1.2
Access Protocols for Star
The OSI model [4] for computer communication defines a hierarchical layering of com-munication protocols. Each layer performs a set of network functions which are used by the layers above this layer. The lowest layer is the Physical layer to the physical medium. The Data Link layer and Network layer are the next higher layers. The Data Link layer is subdivided into Media Access and Logical Link Control layers. The Network layer is subdivided to provide two main functions: Routing and Congestion Control. The Transport layer lies above the Network layer and uses the services offered by the Network layer.
Media access protocols provide access arbitration for the network nodes to the trans-mission medium. Access schemes vary according to the network interconnection topology. Some common topologies used for local area networks are star, bus and ring. In the star topology, all nodes transmit to a star coupler which transmits data back to all the nodes. In the bus topology, nodes are connected by pair of unidirectional buses carrying data. Their is a linear ordering of the nodes based on the node positioning on the bus. The ring topology interconnects nodes in a circular fashion.
Power budget considerations for fiber optic networks [5] showed that the system size supported by the bus was very small whereas the star could support hundreds of nodes. This led to research interest in the design of access protocols for star networks [6–13].
WDM star networks can be configured as single-hop networks where each node can communicate directly with other nodes. Multi-hop networks impose a virtual topology such as Shufflenet on the physical star topology [14]. Multi-hop networks usually have a cost advantage over single-hop networks since the former does not require wavelength
tunable devices. Single-hop networks has the potential advantages of optical self-routing and all-optical communication between source and destination. This dissertation focuses on single-hop networks.
Media access protocols developed for single-hop WDM star networks have been clas-sified into reservation and pre-allocation based protocols. Reservation techniques require at least one of the channels to operate as a control channel which is used to reserve access to the remaining channels (designated as data channels). Pre-allocation strategies operate without a separate control channel – all channels are used for data transmission. The latter approach appears promising due to its low implementation complexity and potential low cost. The first part of this dissertation studies the design and analysis of pre-allocation access protocols for the WDM star.
1.3
Access Protocols for Hierarchical Architecture
The star topology could perhaps support up to hundreds of nodes without any form of amplification. The goal is to provide communication for networks consisting of thousands of nodes. This can be achieved in two possible ways: use optical amplification, or build a hierarchical network from smaller sized clusters. Typical networks would combine the advantages of both these approaches to increase system size.
The improvements of optical amplifiers [2,15] have the potential of extending the system size of bus based networks, motivating their reconsideration. The bus with amplification has the capacity to support hundreds of nodes [2, 16]. However, the star was still shown to retain its excellent fanout relative to the bus with optical amplifiers. In addition to providing implicit node ordering, the bus is characterized by lower fiber but higher coupler requirements [17], and requires amplifiers even for a reasonable network size. The star,
on the other hand, is characterized by large fanout and high fault tolerance, but requires more fiber. The second part of this dissertation evaluates the performance of bus and star topologies for fiber optic networks with optical amplifiers.
A scalable, hierarchical all-optical architecture has been developed to achieve the sys-tem objectives of low-latency and low-cost through a combination of spatial reuse of WDM channels [18]. The intention is to provide a reconfigurable structure that supports distributed shared memory (DSM) [19] and capitalize on any reference locality above uniform [20]. Each processor is associated with a portion of the system address space which can be accessed by other processors through message passing mechanisms. However, the com-munication appears to the user and operating system as a shared memory reference. Cache coherence protocols (snooping or directory based) have to be provided to maintain cache consistency across all the processors [21]. The performance of the DSM organization and the cache coherence protocol is dependent on the media access protocol. Careful attention has to be devoted to the design of the access schemes in an effort to minimize latency and deliver the performance advantage to the application level.
The objective of this hierarchical architecture is to achieve scalability yet avoid the requirement of multiple wavelength tunable devices per node. Furthermore, single-hop communication is achieved: a packet remains in the optical form from source to destination and does not require intermediate routing. This all-optical characteristics is achieved for both data transport and control.
Access protocols that support the DSM organization and associated cache coherence mechanisms is studied as the third part of this dissertation. The protocol studied in this dissertation, called FatMAC, is a hybrid approach that combines the advantages of receiver pre-allocation and reservation access strategies. The protocol has the following charac-teristics: low implementational complexity, collisionless transmission, low-latency at low
loads and stability at high loads, support of varying propagation delays through interleaving cycles, and support of variable packet sizes without segmentation. This protocol reserves access on a pre-allocated channel through control packets.
The dissertation is organized as follows. Chapter 2 provides background information on the OSI model and WDM networks. Chapter 3 presents the media access protocols studied for a WDM star network. Chapter 4 studies the performance of star and bus topologies for a WDM network with optical amplification. Chapter 5 presents the hierarchical architecture and evaluates the performance of media access protocols for this network. Chapter 6 summarizes the research and identifies future scope of this work.
Photonic Networks
The objective of this chapter is two-fold: first, to present an introduction to fiber optic networks and wavelength division multiplexed (WDM) networks in particular; and second, to discuss media access protocols for multiple access, shared medium WDM networks.
In conventional networks, electronic processing is much faster than the communication network. The conventional networks were characterized by high bit error rates and low bandwidth. Communications bandwidth was a severe technological bottleneck. Protocol processing was elaborate and used redundant error checking at different levels to ensure error-free communication. Protocol and network design incorporated efficient bandwidth utilization as the main objective. In this process, communication latency was usually traded off to conserve bandwidth which was the crucial network resource.
With the introduction of optics for communication, there now exists an enormous imbalance in the technologies of communications and computing. The limitations of electronics place an upper limit on the processing speed - up to two gigabits per second with silicon and up to five gigabits per second with gallium arsenide circuitry. Fiber optics is characterized by very high bandwidth of the order of 25000 Gigabits of capacity at three
transmission bands (0.85, 1.3 and 1.55
m
) [2]. Fiber optic networks are also characterized by bit error rates as low was 10?15. The bottleneck in the network has now shifted to the electronic processing. Protocol and network design have to be reconsidered with reduced latency as one of the main goals. Bandwidth can be traded off to ensure low latency communication.
The classical approach of optic network usage involves replacing the communication medium in existing networks using fiber. The non-classical approach involves redesign of existing networks using the special characteristics of optic networks. One such characteristic is the simultaneous data transmission on different wavelengths on a single fiber. This technique is called wavelength division multiplexing (WDM). This is the focus of the first part of the chapter.
The particular emphasis in this dissertation is on all-optical networks where the informa-tion is maintained throughout its route in the optical domain. The performance limitainforma-tions imposed by the electronic processing are observed only at the two end-points of the net-works. The first step toward utilizing the properties of optical networks is the design of protocols for accessing the network in a shared environment. This is the focus of the second part of this chapter and the rest of this dissertation. The main principles in redesigning protocols for the high bandwidth, low error-rate photonic environment can be summarized as follows:
.
Utilize the low error rate and high bandwidth of optics.
Reduce the speed mismatch between electronics and optics.
Maintain all-optical communication.
Achieve low-latency responseProtocol design for computer communication is typically based on a layered approach. The different layers of the protocol model provide different functionalities. Section 2.1 provides a brief introduction to the layered approach in protocol design proposed by the Open Systems Interconnection model. This dissertation is concerned with the development of protocols at the media access level for a photonic network. Section 2.2 provides the background to WDM networks and devices. Section 2.3 discusses previous research and classification of access protocols for WDM networks.
2.1
OSI Model
This section briefly describes the Open Systems Interconnection (OSI) model for computer networks which was approved in 1983 as an international standard by the International Organization for Standardization (ISO) and by the CCITT.
There are two requirements to data communication between two hosts: (i) data delivered by an end user should arrive at the destination correctly and in timely fashion; (ii) data ultimately delivered to the end user is recognizable and in proper form for its correct use. A number of network protocols handle the first part of the problem and higher-level protocols solve the second part.
In a typical network, the higher-level protocols are executed only at the end users. The intermediate nodes execute the network protocols. These two protocols are further broken down into a series of sub-layers leading to the OSI Reference Model as shown in Fig. 2.1. The key concept behind layered communication is the hierarchy between the layers and the fact that each lower layer provides a set of services to the layer above it.
The physical layer and data link layer ensure error-free communication along a single link between two nodes in the network. The function of the physical layer is to ensure that
End user Application Presentation Session Transport Network Physical Data Link Physical Data Link Network Application Presentation Session Transport Network Physical Logical Link Media Access Physical Medium Intermediate Node Higher Layers Lower
Layers Media Access Logical Link
Data Link
End user
Figure 2.1: OSI model for computer communication.
a bit entering the transmission medium at one end reaches the other end without error. The purpose of the data link layer is to ensure that data blocks are reliably transmitted over the network using the underlying physical layer services. The data link layer is further layered into two sub-layers: Media Access Control (MAC) Layer and Logical Link Control (LLC) Layer. The Media Access protocol provides a mechanism for the nodes to gain access to the transmission medium. The Logical Link protocol ensures orderly, correct delivery of packets between neighboring nodes in a network.
The network layer uses the data link layer services and provides a two-fold service: routing data through the network and providing flow/congestion control to prevent network resources from being used up. Routing provides a path or multiple paths that must be set up to connect a source to a destination. Congestion control provides mechanisms needed to guard against statistical fluctuations in traffic which may deplete network resources (such as nodal buffers and transmission links) resulting in congestion. These three layers are grouped and referred to as the network-level protocols and are executed at each intermediate node
of the data transmission path from source to destination.
The transport layer is the lowest of the higher-level protocols and the first end-to-end layer. This is the layer that ensures reliable, sequenced exchange of data between the two end users. It provides the necessary and reliable end-to-end data transmission service for the session layer and other layers above it. It uses the services of the network layer but provides a transparent service so that the details of the network layer and below are hidden from the transport service users. The session layer and other higher level protocols provide user-oriented services and are not of importance to this proposal and will not be discussed any further. The most commonly used transport and network layer protocols are the TCP/IP protocol suite [22].
The layering of protocols as described above imposes an enormous overhead on com-puter communication. A packet needs to be handled by least three different protocol layers. This results in huge overhead due to information exchange between functions operating at different layers and leads to high buffering requirements. Some of the error checking and flow control functions are done at more than one layer. This may not suitable for a high speed environment where efforts are directed towards reduced protocol processing and elimination of the opto-electronic speed mismatch.
One possible solution to this problem is to compact the layers and reduce protocol execution. For example, the data link layer and network layer can be compressed to a single layer which handles media access, link control, flow control and congestion control. Similarly, routing through intermediate nodes can be eliminated by providing virtual point-to-point links between all pairs of nodes. The result will be a compacted layer structure that provides faster communication.
This dissertation studies low complexity protocols for the Media Access sublayer of the OSI model. The next section provides the background for multi-wavelength communication
using wavelength division multiplexing.
2.2
WDM Networks
Wavelength-division (WDM) multiplexing has become a popular choice to circumvent the electronic/optic speed mismatch by partitioning the traffic. Multiple channels, each operating at a rate that the interface electronics can sustain, are formed which eliminate the need that every node receive and process all network traffic as is usually the case with a single channel [1]. Each channel can be switched independently and operate at speeds closer to the electronic interface. WDM networks are established using tunable transmitters and/or receivers to switch between the multiple channels created on the single optical fiber. This dissertation considers the case of both single channel and multiple channel Broadcast-Select WDM networks [1] operating in a multiple access environment.
2.2.1
Optical Components
WDM networks are established using wavelength tunable optical devices. The next sections briefly discuss the principles of wavelength tunable receivers and transmitters. This is followed by a discussion of optical fiber amplifiers including erbium doped fiber amplifiers (EDFA). A complete discussion of this topic may be found in [2].
Tunable Receivers
Wavelength selectivity can be achieved with either a coherent receiver, or a tunable receiver with direct detection. The first approach is more expensive, but has higher channel selec-tivity: channels can be placed closer together so a greater number of channels are formed in the tunable range. A lower cost alternative is a tunable filter with direct detection.
The requirements for a tunable filter are [2]: (a) tunable range or the number of channels it can tune to, and the frequency selectivity of the response after tuning, (b) time required to tune from one channel to another, (c) attenuation, (d) controllability, (e) cost, (f) polarization independence and (g) size and power consumption.
The filters are typically based on interference effects that are wavelength selective. Non-coherent, wavelength tunable filters can be constructed with a variety of techniques: Fabry-Perot and Mach-Zehnder approaches of wavelength dependence of interferometric phenomena, wavelength dependence of coupling through acousto-optic or electro-optic techniques, and resonant amplification that provides gain as well as wavelength selectivity [23, 24]. Fabry-Perot and Mach-Zehnder filters have been constructed where 30 channels have been separated with milli-second switching speed [1]. Active filters have either electro-optic or acousto-optic control. A filter bandwidth of 1
nm
has been achieved with both approaches [23]. However, acousto-optic devices have a tuning range across the full 1.3 - 1.56m
range, while the electro-optic devices are limited to about 15nm
[23, 25]. Acousto-optic devices have switching speeds (s
) slower than electro-optic devices (ns
). Acousto-optic and electro-optic Filters based on polarization beam splitters (PBS) have also been investigated [26–28]. The acousto-optic PBS devices have large tuning ranges and have tuning time in the range ofs
. The electro-optic PBS devices have similar range but have much smaller tuning times (0:
1s
). Tunable filters based on switched-grating approach have been studied in [29–32].Acousto-optic tunable filters (AOTF) for WDM have been described in [24, 25, 28, 33]. AOTFs have advantages of broad tuning range, fast electronic tunability, narrow filter bandwidth, and multiple wavelength selectivity which make them preferable for WDM networks. An optical cross-connect based on AOTFs is described in [1, 25] and can be used to route signals on different wavelength channels independently among different optical
fibers in a WDM system.
Tunable Transmitters
Tunable laser diodes have been an area of active investigation and a summary may be found in [2, 34]. The ideal device is a single-frequency laser diode that can span up to the spectral width of 200
nm
at 1.3 or 1.5m
and is rapidly tunable. The different forms of tunability can be listed as: (a) overall tuning range, (b) continuous tuning range, which is the width of the largest tuning range, and (c) range of continuous control, the range over which control currents and voltages vary monotonically without jumps.Tunable lasers can be achieved through thermal, mechanical, injection-current and acousto-optic means [23, 35]. Thermal and mechanical methods achieve slow tunable devices (
ms
–s
). Injection-current techniques are capable of tuning speeds of a fewns
. External-cavity tunable lasers use an anti-reflection coating on one facet of a Fabry-Perot laser diode and using one of the tunable filters discussed above, between the active laser region and some external mirror reflection. These lasers have tuning-speed problems due to the speed of the tunable filter and the delay for the laser active region to see the new filter setting. WDM networks have become possible through recent advances in narrow linewidth Distributed Feedback (DFB) and Distributed Bragg Reflector (DBR) tunable lasers and filters [34, 36]. The range of a DFB laser has a limit of 10-15nm
due to heating and non-radiative combination [37]. A range of 30nm
has been achieved using a four-section tunable laser in [36]. Phase matched tunable laser diodes use a scheme by which the emitted wavelength is made to depend on the difference in the indices of the two current-injected regions rather than on the index. A tuning range of 57nm
has been achieved using this technique [38].Spectral slicing is a low cost alternative to tunable laser diodes that has been recently in-troduced for this environment [39–43]. A tunable transmitter with
C
channels is constructed usingC
LEDs and a WDM multiplexer. The multiplexer is used to extract the desired wave-length for each channel and block the remaining spectrum of the LED. A system based on spectral splicing has been implemented using off-the-shelf components [42]. Spectral slicing is suitable for short-distance WDM applications when the number of wavelengths required is small.Wavelength control based on the property that the wavelength of the gain peak of Er3+ -doped fiber amplifier decreases as the degree of inversion increases in the fiber. Wavelength tuning over a range of 8
nm
has been achieved using this technique [44]. A new intracavity tuning mechanism based on electro-optic reflection Mach-Zehnder reflectometer provides high speed tuning and wide tuning range of 40nm
[45]. Wavelength tuning with a single control current in a tunable twin-guide DFB laser and a tuning range of 57nm
has been reported in [46].An array of surface emitting laser diodes fabricated on the same wafer can provide a tunable transmitter [47–49]. Tunability is achieved by tuning only one wavelength at a time. The tuning time is determined by the speed of switching from one laser to another.
Optical Amplifiers
Photonic amplification is essentially achieved using two practical approaches: laser diode amplifiers and fiber doped amplifiers. The latter approach is more popular because of the simplicity of manufacture and fiber coupling, polarization independence, wide bandwidth and comparative freedom from crosstalk. Laser diode amplifiers are more expensive to manufacture and have high a level of crosstalk. However, fiber doped amplifiers operate only in the 1.5
m
band while laser diode amplifiers operate equally well in both the 1:
3and 1.5
m
region.Amplifiers can play three basic roles in lightwave systems [2]:
.
The power amplifier to boost the output from a transmitter.
The line amplifier placed between the transmitter and the receiver along the fiber.
The receiver preamplifier to boost the incoming signal before receiver decoding Amplifier placement has a significant impact on the system power budget and has been studied in [2].Two types of laser diode amplifiers are possible: Fabry-Perot amplifiers and traveling wave amplifiers. For network purposes, the traveling-wave amplifiers are more interesting because of their flatter overall gain spectrum and ease of control. Fiber doper amplifiers typically use erbium ions co-doped with alumina. The overall 3 dB bandwidth at 1.5
m
is over 35nm
. For the power amplifier operation, power output values as high as 500 mW has been achieved. Values as high as 46 dB total gain and 10.2 dB gain increase per milliwatt increase of pump power have been reported.Semiconductor amplifiers are very promising for future optical communication systems because of their high gain, broad amplification bandwidth, low power consumption, and compactness. A 4-channel array at 1.3
m
with 20 dB gain and less than 1 dB polarization sensitivity has been reported in [50].Pr3+
doped fiber amplifier is a candidate for use as an optical amplifier in 1.3
m
transmission. This has been indicated by demonstration of a laser diode pumped PDFA with a gain of 28.3 dB. The main issue for improving PDFA performance is to increase the gain coefficient as experimented in [51, 52].This section summarized the characteristics of the optical devices required to establish WDM systems. The next section studies the optical multiple access topologies for a WDM
network.
2.2.2
Network Topology
This section studies the topology of the network to interconnect the nodes. A variety of optical interconnection topologies can be used to achieve a multiple access environment for the photonic network as shown in Fig. 2.2 [53]. The commonly studied topologies are the bus and the star. The bus topology has been further classified as singly folded unidirectional bus, doubly folded unidirectional bus, and dual bus. The bus topology has been favored for local network interconnection since degradation due to propagation delay can be reduced by taking advantage of the physical ordering [53]. The unidirectional nature of the photonic bus establishes a physical ordering among the nodes which is exploited by the Demand Assignment Media Access (DAMA) protocols [53].
Fig. 2.2(a) shows the singly folded unidirectional bus configuration (FUB). The FUB has two unidirectional sections – the outbound bus for data transmission and the inbound bus for data reception. This configuration has been used in protocols including D-Net [54]. Fig. 2.2(b) shows the doubly-folded unidirectional bus (DFUB). This configuration has been used in Expressnet [55]. Both FUB and DFUB configurations require a sense tap and a transmitter on the outbound bus and a receiver on the inbound bus. A dual-bus configuration (DUB) is shown in Fig. 2.2(c) with two independent buses carrying traffic in opposite directions. This approach has been used by Fasnet [56] and DQDB (IEEE 802.6). Each node has a sense tap, transmitter and a receiver on each bus doubling system cost. In a single channel system, the sense tap and receiver functions can be implemented using a single device [56]. In multi-channel systems where devices can be fixed or tunable, the sense tap and receiver may be implemented separately. The physical layout of the star is shown in Fig. 2.2(d). Each node has an outbound fiber to and an inbound fiber from the
T R T R Node m 1 T R Node m M-1 S S S S R T R T R T R R T R Node m 0 R T R T Node m M-1 T R Node m 1
. . .
(c) . . . S S Inbound Outbound S Node m 0 STAR . . .Node m0 Node m1 Node mM-1
S S Bus A Bus B T T Outbound Inbound . . .
Node m0 Node m1 Node mM-1
S T R S T R S T R (a) (b) (d)
Figure 2.2: Optical multiple access channel configurations: (a) folded unidirectional bus, (b) doubly folded unidirectional bus, (c) dual unidirectional bus, and (d) star-coupled configuration.
star.
The optical power budget (OPB) characteristics determine the number of nodes that can be connected in a network while maintaining a specified bit-error-rate. The number of nodes attached to a multiple access channel is bound by the saturation traffic and the optical power budget. The high bandwidth of fiber precludes the saturation limitation and the principal factor of optical fanout is the power budget. Power budget considerations for fiber optic networks [5] showed that the system size supported by the bus was very small whereas the star could support hundreds of nodes. The system sizes of bus and star networks with erbium doped fiber amplifiers (EDFA) have been studied in [2, 16] and the star was also shown to have better fanout for multi-channel systems. A FUB network with an optimized number of amplifiers (Gain of 15 dB) and active reciprocal couplers was shown to have a fanout of over 400 nodes, whereas a star network with receiver pre-amplification was shown to support over 2000 nodes [2].
The star topology exhibits the following characteristics which make it a better alternative the bus topology:
1. System size limitations in terms of the optical power budget were compared in [5, 57] and the star-coupled configuration was shown to exhibit superior fanout characteris-tics over bus-based topologies. However, bus based networks have renewed interest due to recent developments on optic amplifiers.
2. Star-coupled networks have high fault-tolerance due to their passive nature and complete unity distance connectivity [57, 58].
This section provided a background on WDM networks including tunable transmitters, receivers and optical amplifiers. The following section examines the WDM media access mechanisms for the star topology.
2.3
WDM Media Access
This section discusses classification of WDM networks. This is followed by a discussion of multiple access protocols for a WDM star coupled network.
2.3.1
Network Classification
WDM Networks have been classified using a variety of criterion based on their character-istics. Three classes of WDM networks are Wavelength Routing (WR), Broadcast-Select (BS) and Wavelength Switching (WS) [1]. In a WR network, the transmitted wavelength completely determines the path in the network. This needs wavelength selectable devices. A Broadcast-Select network is passive with no internal wavelength selectivity. Selectivity is achieved by using tunable transmitters and/or tunable receivers. BS network can be configured to provide optical multiple access channels [59, 60], or virtual point-to-point interconnection in the wavelength domain [14, 61–71]. Wavelength switching can be done in two ways: dynamically where WDM routing is changed in the network to switch from one path to another, or by wavelength conversion where a signal is transferred from one to another on an optical fiber. WS devices based on AOTFs [25, 72] or wavelength convert-ers [73,74] could be used to realize WS networks. WMCH [75] and FHA [18] architectures are based on BS passive star-coupled configuration.
This proposal considers a BS network operating in a multi-access environment. As mentioned above, BS networks can be classified as single-hop multiple access [1] or multi-hop virtual topology networks [61].
Single-hop networks: provide a direct path between source and destination without
inter-mediate routing. Packets are transmitted on an all-optical path from source to destination. Multiple nodes can simultaneously access the medium. Single-hop networks require at
least one tunable device to accomplish transmission. Their main advantage is the simplicity of node addition/deletion and the elimination of intermediate routing and electronic pro-cessing. Single-hop networks typically require at least one wavelength tunable device per node.
Multi-hop networks: represent regular or irregular graph topologies implement ed in the
wavelength domain. In this case, a “virtual” topology is embedded on the physical star topology [61]. They require intermediate routing nodes and hence increasing packet latency due to electronic processing [14]. Multi-hop networks typically do not require wavelength tunable devices and have a potential cost-advantage over single-hop networks. However, the number of channels may be linked to the number of nodes limiting the system size [14]. It is possible to embed any virtual topology on the physical star topology. This is achieved by setting the wavelengths of the transmitters and receivers of the nodes. For example, consider a set of 3 nodes interconnected using a physical star topology. Let each node have two transmitters and two receivers; and let f
ij1i
6g represent the set of available wavelengths. The following wavelength allocation establishes a virtual ring topology:Node
T
1T
2R
1R
21
1 2 4 52
3 4 1 61
5 6 2 3The perfect shuffle implementation on a star topology is shown in Fig 2.3 and has been investigated in [14]. The performance of a multi-hop network is determined by its virtual topology. A summary of access protocols for single-hop and multiple-hop networks may be found in [76, 77].
λ1 λ2 λ3 λ4 λ5 λ6 λ7 λ8 λ1 λ5 λ2 λ6 λ3 λ7 λ4 λ8 λ9 λ10 λ11 λ12 λ13 λ14 λ15 λ16 λ13 λ9 λ10 λ14 λ11 λ15 λ12 λ16
M=8 nodes
C=16 channels
Topology emulated in
wavelength domain
Physical topology is star
4
3
2
1
5
6
7
8
4
3
2
1
assignment
through transmitter/receiver
Logical connection
Figure 2.3: Shufflenet implementation on the physical star topology.
Depending on the architecture, optical self-routing can be achieved where a node receives only data destined to it and the system has non-blocking crossbar connectivity [57]. Optical self-routing partitions the traffic relaxing the design constraints on the receiver subsystem. This is crucial since the photonic network has higher throughput support than the packet processing capabilities.
This dissertation focuses on photonic star-coupled passive networks that use WDM to create wavelength division multiple access (WDMA) channels. The reasons for choosing this approach are:
.
Provide an all-optical network.
higher topological connectivity.
low system complexity.
avoid intermediate processing and delayThe next section studies multiple access protocols for a single channel system with broadcast. This is followed by a discussion of protocols studied for the multi-channel environment.
2.3.2
Single Channel Multiple Access
Multiple access protocols for single channel broadcast networks have been extensively studied. A summary of multi-access protocols for these networks may be found in [78– 81]. The protocols are typically classified as contention based and non contention based protocols. Random access or contention based protocols allow more than one node to access the medium at the same time resulting in possible collisions. The protocol usually defines a contention resolution mechanism in the event of a collision. Static access or non-contention based protocols pre-determine the time instants a node may access the network avoiding collisions. The simplest random access protocols are Aloha (A) and Slotted Aloha (SA) [81]. Time Division Multiplexing (TDM) is a typical example of a static access protocol [81].
Performance evaluation of access schemes are usually based on packet delay (latency), channel utilization, and access fairness. Packet delay is the average time between packet generation at the source and reception at the destination. Channel utilization is the fraction of time the network is utilized for useful data transmission. Network throughput is defined as the number of packets that are sent successfully per unit time. Access fairness defines whether all nodes share the medium equally without undue advantage to any particular set of nodes. The performance of protocols is typically studied with varying number of nodes and the system traffic load.
N/2 1 C Allocation Static Access Random THROUGHPUT DELAY
Figure 2.4: Performance characterization of Random and Static Access protocols.
packet delay under light loads, (b) potential instability and low utilization under heavy traffic, (c) possible unfair medium sharing. The performance of static access schemes can be characterized as: (a) high packet delay under light traffic, (b) stable performance and high utilization under heavy traffic, and (c) fair access to all nodes. Fig. 2.4 summarizes the delay-throughput of the two approaches. Other extensions to these two approaches include reservation-SA, reservation-TDMA, Carrier Sense Multiple Access (CSMA), contention tree protocol [81]. These protocols have performance characteristics that lie between these SA and TDM.
The next section examines extension of these principles to multiple channel WDM networks.
2.3.3
WDM Multiple Access
Multiple access protocols for WDM networks can be classified into two major categories: Reservation Based and Pre-Allocated Based strategies.
Reservation Based: Reservation strategies designate one of the
C
channels as the control channel which is used to coordinate data transmission between the nodes. The remaining channels are used for data transmission and are referred to as data channels. Each node typically has (i) a transmitter and receiver which are tuned to the control channel; and (ii) a transmitter/receiver which can tune to the data channels.For example, if node
A
has to send a packet to nodeB
, it transmits a control packet on the control channel which is monitored by all receivers. The control packet contains the destination node and the data channel in the control packet. Conflicts can occur when more than one node attempts transmission at the same time on the control channel or the same data channel. The media access protocol ensures that the control packet is transmitted successfully. NodeB
, on receiving this control packet, sets up its receiver to receive the data packet on the channel indicated in the control packet. Media access protocols are now required to provide arbitration and resolve conflicts on the data channels too.A typical control channel based MAC protocol for the multiple channel WDM network is designated as X/Y, where X and Y are implemented as the MAC protocols for the control and data channels respectively. Some combinations of such protocols introduced in [6] are A/A, SA/A, A/CSMA, CSMA/A. A discussion of control channel protocols for WDM networks may be found in [76].
Pre-Allocation Based: Unlike reservation based schemes, pre-allocation based protocols
do not require a separate control channel. All channels are used for data transmission. This approach requires either a tunable receiver or a tunable transmitter. If the transmitter is tunable, the receiver is fixed and receives data only on one channel (home channel ). A transmitter tunes to the home channel of the receiver and transmits the packet. If
M > C
, then a home channel is shared among more than one node. IfM
=C
, each node has its own channel and self-routing is achieved. If the receiver is tunable, the transmitter is fixed on onechannel (home channel) and the receiver tunes to the transmitter’s home channel. The latter approach requires coordination for the receiver to tune to the transmitter’s home channel. We have investigated the first approach because it is simpler and tunable transmitters are less expensive than tunable receivers.
The idea of pre-allocation was initially considered in [9, 82, 83]. The reservation approach has attracted a lot of research attention as presented in [6, 76]. Reservation protocols promise greater flexibility in the availability of data channels. However, some protocols proposed link the number of channels to the number of nodes in the system [8,12]. The number of WDM channels that may be available theoretically may be as high as 100. A more realistic number determined by device characteristics limits the number of channels between 4 to 16. Let
M
denote the number of nodes in the network andC
denote the number of multiple-access channels. The protocol design in this dissertation is targeted towards a system with many more nodes than the number of channels (M
C
) for small values ofC
. A pre-allocation approach that does not utilize a control channel is considered here. This approach appears to be very promising due to its low implementational and operational complexity.The simplest configuration in terms of system cost uses one tunable element and one fixed element. The protocols studied in this thesis are based on one tunable receiver and fixed receiver per node. The various choices for this configuration are examined in the next section.
2.3.4
Node Architecture
Each node in a WDM network requires at least one tunable transmitter or tunable receiver. A node can have more transmitter/receiver pairs that can be tunable or fixed. The particular configuration is determined chiefly by the access protocol. Fig. 2.5 shows a set of possible
F T F T F F F F T F F F T T T F T T F
Node 1 Node 2 Node M
F
T T
F
Node 1 Node 2 Node M
F
T T
F
Node 1 Node 2 Node M
(a)
(b)
(c)
network architectures.
Fig. 2.5(a) shows a system where each node has a single transmitter and a receiver. There are three possible configurations with this organization: tunable transmitter and fixed receiver, fixed transmitter and tunable receiver, and tunable transmitter/receiver. One particular configuration uses a tunable transmitter and one fixed receiver [84]. Each node listens to one channel referred to as its home channel. The home channel allocation to the nodes can be determined based on a uniform policy or based on traffic requirements. This approach has potential for low cost since tunable receivers are more expensive than tunable transmitters or fixed receivers.
Fig. 2.5(b) shows a system with a tunable transmitter and two receivers. There are various choices depending on the tunability of the receivers. One particular configuration uses one fixed receiver and one tunable receiver per node as studied in [10]. The protocol is based on a control channel. The fixed receiver remains tuned to the control channel. The transmitter first sends a control packet on the control channel followed by data on one of the other channels.
Fig. 2.5(c) shows a system with a pair of transmitters and receivers. One possible configuration uses two fixed transmitters, one fixed receiver and one tunable receiver per node. The protocol based on this system [8] allocates uses one control channel. Each node is assigned a data channel on which it transmits data. The fixed receiver is tuned to the control channel. A backlogged node transmits a control packet in its slot in the next control cycle and then sends its packet on its own data channel. The receiver has to tune to that channel to receive the packet.
The above discussion presented the various choices for the node architecture. This dissertation examines protocols for the architecture shown in Fig. 2.5(a) with a single tunable transmitter and fixed receiver per node. This architecture is also used in a hierarchical
network architecture used in a distributed shared memory (DSM) optically interconnected multiprocessor system [18, 85].
2.4
Summary
This chapter provided a background on WDM photonic networks. The OSI model of computer communication was presented and the significance of efficient media access protocols considered. Previous research on access protocols for WDM star networks and their classification was presented. The next chapter presents media access protocols studied for this dissertation for a passive-star coupled network.
Access Protocols for WDM Star
Networks
In this chapter, access protocols for a passive star-coupled WDM network are examined. A multiple access environment can be achieved through a variety of optical channel topologies [53] as shown in Fig. 2.2. Two main topologies studied for local networks are the bus and star. This chapter examines WDM star networks without photonic amplification. The star has excellent fanout and fault tolerance characteristics compared to the bus which forms the chief motivation for examining protocols for a passive star coupled WDM network.
Section 3.1 discusses the factors affecting protocol design and the impact of reserva-tion and pre-allocareserva-tion approaches on design and performance. Secreserva-tion 3.2 provides the description of the pre-allocation protocols studied as part of this dissertation. Section 3.3 provides the performance analysis of the protocols based on semi-Markov analytic models and discrete-event simulation models.
3.1
Factors affecting protocol design
Access protocols for the star have been classified as reservation and pre-allocation proto-cols as discussed in Section 2.3.3. The source node in a reservation protocol uses some form of pre-transmission coordination to determine the channel of communication. The pre-allocation protocols do not require such coordination and allocate channels to either receivers or transmitter. The factors identified below are used to examine the suitability of the pre-allocation protocols. In general, the problems are not new but have taken on increased importance due to the relation of their magnitude to the nominal data packet transmission time in high-speed networks:
Propagation delay: is the time required for light energy to travel from source to destina-tion. It has been shown that large ratios of transmission time to propagation delay result in degraded performance [53]. Protocol design has to consider the impact of propagation delay on protocol performance and implementation.
For example, light requires approximately 30
ms
to propagate from New York to Los Angeles assuming light is traveling at about two-thirds of its maximum speed. Transmission of a 1Mb packet requires only 1ms
with a transmission rate of 1 Gbps. For smaller distances, such as with a LAN, this concern still holds since the unit of transfer is also smaller (of the order of kilobits).Reservation based schemes require control packets which are typically much smaller than data packets. The ratio of control packet transmission time to propagation delay is larger than that for data packets resulting in degraded performance. This has motivated the investigation of pre-allocation protocols which do not require control packets.
Processing Latency: is defined as the processing time preceding packet transmission. It has two main components of delay – the tuning latency of the optical devices and protocol processing overhead.
1. Tuning Latency is defined as the time spent in tuning the wavelength tunable trans-mitters and receivers(filters). Tunable transtrans-mitters with tuning speeds of 15
ns
and tuning range of 50 channels have been demonstrated [37]. Tunable filters have been shown to have tuning speeds varying froms
tons
[25].The tuning latency of these optical devices is not negligible and has a significant impact on overall system performance [86, 87]. In general, faster tuning requires higher cost. An objective of this work is to achieve more cost effective network solutions by developing techniques that perform well with slower devices by hiding the processing latencies.
The proposed protocols use a tunable transmitter and a fixed receiver. The tunable transmitter may be replaced by a laser array transmitter where each laser is always tuned to a single wavelength. This reduces the impact of transmitter tuning time between channels.
2. Protocol processing overhead is incurred for both incoming and outgoing packets. Incoming packets require opto-electronic conversion, header decoding and checksum verification. Outgoing packets require time to execute the protocol algorithm, to select the packet for transmission and to construct and append the header. The protocols should be simple to reduce the time spent in processing which might include buffer manipulation, arbitration algorithms, status table updates, etc.
Control channel based schemes require all nodes to process all control channel traffic. This places a very high demand on the receiver subsystem, in terms of its throughput